8 :Author: Tejun Heo <tj@kernel.org>
10 This is the authoritative documentation on the design, interface and
11 conventions of cgroup v2. It describes all userland-visible aspects
12 of cgroup including core and specific controller behaviors. All
13 future changes must be reflected in this document. Documentation for
14 v1 is available under :ref:`Documentation/admin-guide/cgroup-v1/index.rst <cgroup-v1>`.
23 2-2. Organizing Processes and Threads
26 2-3. [Un]populated Notification
27 2-4. Controlling Controllers
28 2-4-1. Enabling and Disabling
29 2-4-2. Top-down Constraint
30 2-4-3. No Internal Process Constraint
32 2-5-1. Model of Delegation
33 2-5-2. Delegation Containment
35 2-6-1. Organize Once and Control
36 2-6-2. Avoid Name Collisions
37 3. Resource Distribution Models
45 4-3. Core Interface Files
48 5-1-1. CPU Interface Files
50 5-2-1. Memory Interface Files
51 5-2-2. Usage Guidelines
52 5-2-3. Memory Ownership
54 5-3-1. IO Interface Files
57 5-3-3-1. How IO Latency Throttling Works
58 5-3-3-2. IO Latency Interface Files
61 5-4-1. PID Interface Files
63 5.5-1. Cpuset Interface Files
66 5-7-1. RDMA Interface Files
68 5.8-1. HugeTLB Interface Files
70 5.9-1 Miscellaneous cgroup Interface Files
71 5.9-2 Migration and Ownership
74 5-N. Non-normative information
75 5-N-1. CPU controller root cgroup process behaviour
76 5-N-2. IO controller root cgroup process behaviour
79 6-2. The Root and Views
80 6-3. Migration and setns(2)
81 6-4. Interaction with Other Namespaces
82 P. Information on Kernel Programming
83 P-1. Filesystem Support for Writeback
84 D. Deprecated v1 Core Features
85 R. Issues with v1 and Rationales for v2
86 R-1. Multiple Hierarchies
87 R-2. Thread Granularity
88 R-3. Competition Between Inner Nodes and Threads
89 R-4. Other Interface Issues
90 R-5. Controller Issues and Remedies
100 "cgroup" stands for "control group" and is never capitalized. The
101 singular form is used to designate the whole feature and also as a
102 qualifier as in "cgroup controllers". When explicitly referring to
103 multiple individual control groups, the plural form "cgroups" is used.
109 cgroup is a mechanism to organize processes hierarchically and
110 distribute system resources along the hierarchy in a controlled and
113 cgroup is largely composed of two parts - the core and controllers.
114 cgroup core is primarily responsible for hierarchically organizing
115 processes. A cgroup controller is usually responsible for
116 distributing a specific type of system resource along the hierarchy
117 although there are utility controllers which serve purposes other than
118 resource distribution.
120 cgroups form a tree structure and every process in the system belongs
121 to one and only one cgroup. All threads of a process belong to the
122 same cgroup. On creation, all processes are put in the cgroup that
123 the parent process belongs to at the time. A process can be migrated
124 to another cgroup. Migration of a process doesn't affect already
125 existing descendant processes.
127 Following certain structural constraints, controllers may be enabled or
128 disabled selectively on a cgroup. All controller behaviors are
129 hierarchical - if a controller is enabled on a cgroup, it affects all
130 processes which belong to the cgroups consisting the inclusive
131 sub-hierarchy of the cgroup. When a controller is enabled on a nested
132 cgroup, it always restricts the resource distribution further. The
133 restrictions set closer to the root in the hierarchy can not be
134 overridden from further away.
143 Unlike v1, cgroup v2 has only single hierarchy. The cgroup v2
144 hierarchy can be mounted with the following mount command::
146 # mount -t cgroup2 none $MOUNT_POINT
148 cgroup2 filesystem has the magic number 0x63677270 ("cgrp"). All
149 controllers which support v2 and are not bound to a v1 hierarchy are
150 automatically bound to the v2 hierarchy and show up at the root.
151 Controllers which are not in active use in the v2 hierarchy can be
152 bound to other hierarchies. This allows mixing v2 hierarchy with the
153 legacy v1 multiple hierarchies in a fully backward compatible way.
155 A controller can be moved across hierarchies only after the controller
156 is no longer referenced in its current hierarchy. Because per-cgroup
157 controller states are destroyed asynchronously and controllers may
158 have lingering references, a controller may not show up immediately on
159 the v2 hierarchy after the final umount of the previous hierarchy.
160 Similarly, a controller should be fully disabled to be moved out of
161 the unified hierarchy and it may take some time for the disabled
162 controller to become available for other hierarchies; furthermore, due
163 to inter-controller dependencies, other controllers may need to be
166 While useful for development and manual configurations, moving
167 controllers dynamically between the v2 and other hierarchies is
168 strongly discouraged for production use. It is recommended to decide
169 the hierarchies and controller associations before starting using the
170 controllers after system boot.
172 During transition to v2, system management software might still
173 automount the v1 cgroup filesystem and so hijack all controllers
174 during boot, before manual intervention is possible. To make testing
175 and experimenting easier, the kernel parameter cgroup_no_v1= allows
176 disabling controllers in v1 and make them always available in v2.
178 cgroup v2 currently supports the following mount options.
181 Consider cgroup namespaces as delegation boundaries. This
182 option is system wide and can only be set on mount or modified
183 through remount from the init namespace. The mount option is
184 ignored on non-init namespace mounts. Please refer to the
185 Delegation section for details.
188 Reduce the latencies of dynamic cgroup modifications such as
189 task migrations and controller on/offs at the cost of making
190 hot path operations such as forks and exits more expensive.
191 The static usage pattern of creating a cgroup, enabling
192 controllers, and then seeding it with CLONE_INTO_CGROUP is
193 not affected by this option.
196 Only populate memory.events with data for the current cgroup,
197 and not any subtrees. This is legacy behaviour, the default
198 behaviour without this option is to include subtree counts.
199 This option is system wide and can only be set on mount or
200 modified through remount from the init namespace. The mount
201 option is ignored on non-init namespace mounts.
204 Recursively apply memory.min and memory.low protection to
205 entire subtrees, without requiring explicit downward
206 propagation into leaf cgroups. This allows protecting entire
207 subtrees from one another, while retaining free competition
208 within those subtrees. This should have been the default
209 behavior but is a mount-option to avoid regressing setups
210 relying on the original semantics (e.g. specifying bogusly
211 high 'bypass' protection values at higher tree levels).
213 memory_hugetlb_accounting
214 Count HugeTLB memory usage towards the cgroup's overall
215 memory usage for the memory controller (for the purpose of
216 statistics reporting and memory protetion). This is a new
217 behavior that could regress existing setups, so it must be
218 explicitly opted in with this mount option.
220 A few caveats to keep in mind:
222 * There is no HugeTLB pool management involved in the memory
223 controller. The pre-allocated pool does not belong to anyone.
224 Specifically, when a new HugeTLB folio is allocated to
225 the pool, it is not accounted for from the perspective of the
226 memory controller. It is only charged to a cgroup when it is
227 actually used (for e.g at page fault time). Host memory
228 overcommit management has to consider this when configuring
229 hard limits. In general, HugeTLB pool management should be
230 done via other mechanisms (such as the HugeTLB controller).
231 * Failure to charge a HugeTLB folio to the memory controller
232 results in SIGBUS. This could happen even if the HugeTLB pool
233 still has pages available (but the cgroup limit is hit and
234 reclaim attempt fails).
235 * Charging HugeTLB memory towards the memory controller affects
236 memory protection and reclaim dynamics. Any userspace tuning
237 (of low, min limits for e.g) needs to take this into account.
238 * HugeTLB pages utilized while this option is not selected
239 will not be tracked by the memory controller (even if cgroup
240 v2 is remounted later on).
243 Organizing Processes and Threads
244 --------------------------------
249 Initially, only the root cgroup exists to which all processes belong.
250 A child cgroup can be created by creating a sub-directory::
254 A given cgroup may have multiple child cgroups forming a tree
255 structure. Each cgroup has a read-writable interface file
256 "cgroup.procs". When read, it lists the PIDs of all processes which
257 belong to the cgroup one-per-line. The PIDs are not ordered and the
258 same PID may show up more than once if the process got moved to
259 another cgroup and then back or the PID got recycled while reading.
261 A process can be migrated into a cgroup by writing its PID to the
262 target cgroup's "cgroup.procs" file. Only one process can be migrated
263 on a single write(2) call. If a process is composed of multiple
264 threads, writing the PID of any thread migrates all threads of the
267 When a process forks a child process, the new process is born into the
268 cgroup that the forking process belongs to at the time of the
269 operation. After exit, a process stays associated with the cgroup
270 that it belonged to at the time of exit until it's reaped; however, a
271 zombie process does not appear in "cgroup.procs" and thus can't be
272 moved to another cgroup.
274 A cgroup which doesn't have any children or live processes can be
275 destroyed by removing the directory. Note that a cgroup which doesn't
276 have any children and is associated only with zombie processes is
277 considered empty and can be removed::
281 "/proc/$PID/cgroup" lists a process's cgroup membership. If legacy
282 cgroup is in use in the system, this file may contain multiple lines,
283 one for each hierarchy. The entry for cgroup v2 is always in the
286 # cat /proc/842/cgroup
288 0::/test-cgroup/test-cgroup-nested
290 If the process becomes a zombie and the cgroup it was associated with
291 is removed subsequently, " (deleted)" is appended to the path::
293 # cat /proc/842/cgroup
295 0::/test-cgroup/test-cgroup-nested (deleted)
301 cgroup v2 supports thread granularity for a subset of controllers to
302 support use cases requiring hierarchical resource distribution across
303 the threads of a group of processes. By default, all threads of a
304 process belong to the same cgroup, which also serves as the resource
305 domain to host resource consumptions which are not specific to a
306 process or thread. The thread mode allows threads to be spread across
307 a subtree while still maintaining the common resource domain for them.
309 Controllers which support thread mode are called threaded controllers.
310 The ones which don't are called domain controllers.
312 Marking a cgroup threaded makes it join the resource domain of its
313 parent as a threaded cgroup. The parent may be another threaded
314 cgroup whose resource domain is further up in the hierarchy. The root
315 of a threaded subtree, that is, the nearest ancestor which is not
316 threaded, is called threaded domain or thread root interchangeably and
317 serves as the resource domain for the entire subtree.
319 Inside a threaded subtree, threads of a process can be put in
320 different cgroups and are not subject to the no internal process
321 constraint - threaded controllers can be enabled on non-leaf cgroups
322 whether they have threads in them or not.
324 As the threaded domain cgroup hosts all the domain resource
325 consumptions of the subtree, it is considered to have internal
326 resource consumptions whether there are processes in it or not and
327 can't have populated child cgroups which aren't threaded. Because the
328 root cgroup is not subject to no internal process constraint, it can
329 serve both as a threaded domain and a parent to domain cgroups.
331 The current operation mode or type of the cgroup is shown in the
332 "cgroup.type" file which indicates whether the cgroup is a normal
333 domain, a domain which is serving as the domain of a threaded subtree,
334 or a threaded cgroup.
336 On creation, a cgroup is always a domain cgroup and can be made
337 threaded by writing "threaded" to the "cgroup.type" file. The
338 operation is single direction::
340 # echo threaded > cgroup.type
342 Once threaded, the cgroup can't be made a domain again. To enable the
343 thread mode, the following conditions must be met.
345 - As the cgroup will join the parent's resource domain. The parent
346 must either be a valid (threaded) domain or a threaded cgroup.
348 - When the parent is an unthreaded domain, it must not have any domain
349 controllers enabled or populated domain children. The root is
350 exempt from this requirement.
352 Topology-wise, a cgroup can be in an invalid state. Please consider
353 the following topology::
355 A (threaded domain) - B (threaded) - C (domain, just created)
357 C is created as a domain but isn't connected to a parent which can
358 host child domains. C can't be used until it is turned into a
359 threaded cgroup. "cgroup.type" file will report "domain (invalid)" in
360 these cases. Operations which fail due to invalid topology use
361 EOPNOTSUPP as the errno.
363 A domain cgroup is turned into a threaded domain when one of its child
364 cgroup becomes threaded or threaded controllers are enabled in the
365 "cgroup.subtree_control" file while there are processes in the cgroup.
366 A threaded domain reverts to a normal domain when the conditions
369 When read, "cgroup.threads" contains the list of the thread IDs of all
370 threads in the cgroup. Except that the operations are per-thread
371 instead of per-process, "cgroup.threads" has the same format and
372 behaves the same way as "cgroup.procs". While "cgroup.threads" can be
373 written to in any cgroup, as it can only move threads inside the same
374 threaded domain, its operations are confined inside each threaded
377 The threaded domain cgroup serves as the resource domain for the whole
378 subtree, and, while the threads can be scattered across the subtree,
379 all the processes are considered to be in the threaded domain cgroup.
380 "cgroup.procs" in a threaded domain cgroup contains the PIDs of all
381 processes in the subtree and is not readable in the subtree proper.
382 However, "cgroup.procs" can be written to from anywhere in the subtree
383 to migrate all threads of the matching process to the cgroup.
385 Only threaded controllers can be enabled in a threaded subtree. When
386 a threaded controller is enabled inside a threaded subtree, it only
387 accounts for and controls resource consumptions associated with the
388 threads in the cgroup and its descendants. All consumptions which
389 aren't tied to a specific thread belong to the threaded domain cgroup.
391 Because a threaded subtree is exempt from no internal process
392 constraint, a threaded controller must be able to handle competition
393 between threads in a non-leaf cgroup and its child cgroups. Each
394 threaded controller defines how such competitions are handled.
397 [Un]populated Notification
398 --------------------------
400 Each non-root cgroup has a "cgroup.events" file which contains
401 "populated" field indicating whether the cgroup's sub-hierarchy has
402 live processes in it. Its value is 0 if there is no live process in
403 the cgroup and its descendants; otherwise, 1. poll and [id]notify
404 events are triggered when the value changes. This can be used, for
405 example, to start a clean-up operation after all processes of a given
406 sub-hierarchy have exited. The populated state updates and
407 notifications are recursive. Consider the following sub-hierarchy
408 where the numbers in the parentheses represent the numbers of processes
414 A, B and C's "populated" fields would be 1 while D's 0. After the one
415 process in C exits, B and C's "populated" fields would flip to "0" and
416 file modified events will be generated on the "cgroup.events" files of
420 Controlling Controllers
421 -----------------------
423 Enabling and Disabling
424 ~~~~~~~~~~~~~~~~~~~~~~
426 Each cgroup has a "cgroup.controllers" file which lists all
427 controllers available for the cgroup to enable::
429 # cat cgroup.controllers
432 No controller is enabled by default. Controllers can be enabled and
433 disabled by writing to the "cgroup.subtree_control" file::
435 # echo "+cpu +memory -io" > cgroup.subtree_control
437 Only controllers which are listed in "cgroup.controllers" can be
438 enabled. When multiple operations are specified as above, either they
439 all succeed or fail. If multiple operations on the same controller
440 are specified, the last one is effective.
442 Enabling a controller in a cgroup indicates that the distribution of
443 the target resource across its immediate children will be controlled.
444 Consider the following sub-hierarchy. The enabled controllers are
445 listed in parentheses::
447 A(cpu,memory) - B(memory) - C()
450 As A has "cpu" and "memory" enabled, A will control the distribution
451 of CPU cycles and memory to its children, in this case, B. As B has
452 "memory" enabled but not "CPU", C and D will compete freely on CPU
453 cycles but their division of memory available to B will be controlled.
455 As a controller regulates the distribution of the target resource to
456 the cgroup's children, enabling it creates the controller's interface
457 files in the child cgroups. In the above example, enabling "cpu" on B
458 would create the "cpu." prefixed controller interface files in C and
459 D. Likewise, disabling "memory" from B would remove the "memory."
460 prefixed controller interface files from C and D. This means that the
461 controller interface files - anything which doesn't start with
462 "cgroup." are owned by the parent rather than the cgroup itself.
468 Resources are distributed top-down and a cgroup can further distribute
469 a resource only if the resource has been distributed to it from the
470 parent. This means that all non-root "cgroup.subtree_control" files
471 can only contain controllers which are enabled in the parent's
472 "cgroup.subtree_control" file. A controller can be enabled only if
473 the parent has the controller enabled and a controller can't be
474 disabled if one or more children have it enabled.
477 No Internal Process Constraint
478 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
480 Non-root cgroups can distribute domain resources to their children
481 only when they don't have any processes of their own. In other words,
482 only domain cgroups which don't contain any processes can have domain
483 controllers enabled in their "cgroup.subtree_control" files.
485 This guarantees that, when a domain controller is looking at the part
486 of the hierarchy which has it enabled, processes are always only on
487 the leaves. This rules out situations where child cgroups compete
488 against internal processes of the parent.
490 The root cgroup is exempt from this restriction. Root contains
491 processes and anonymous resource consumption which can't be associated
492 with any other cgroups and requires special treatment from most
493 controllers. How resource consumption in the root cgroup is governed
494 is up to each controller (for more information on this topic please
495 refer to the Non-normative information section in the Controllers
498 Note that the restriction doesn't get in the way if there is no
499 enabled controller in the cgroup's "cgroup.subtree_control". This is
500 important as otherwise it wouldn't be possible to create children of a
501 populated cgroup. To control resource distribution of a cgroup, the
502 cgroup must create children and transfer all its processes to the
503 children before enabling controllers in its "cgroup.subtree_control"
513 A cgroup can be delegated in two ways. First, to a less privileged
514 user by granting write access of the directory and its "cgroup.procs",
515 "cgroup.threads" and "cgroup.subtree_control" files to the user.
516 Second, if the "nsdelegate" mount option is set, automatically to a
517 cgroup namespace on namespace creation.
519 Because the resource control interface files in a given directory
520 control the distribution of the parent's resources, the delegatee
521 shouldn't be allowed to write to them. For the first method, this is
522 achieved by not granting access to these files. For the second, the
523 kernel rejects writes to all files other than "cgroup.procs" and
524 "cgroup.subtree_control" on a namespace root from inside the
527 The end results are equivalent for both delegation types. Once
528 delegated, the user can build sub-hierarchy under the directory,
529 organize processes inside it as it sees fit and further distribute the
530 resources it received from the parent. The limits and other settings
531 of all resource controllers are hierarchical and regardless of what
532 happens in the delegated sub-hierarchy, nothing can escape the
533 resource restrictions imposed by the parent.
535 Currently, cgroup doesn't impose any restrictions on the number of
536 cgroups in or nesting depth of a delegated sub-hierarchy; however,
537 this may be limited explicitly in the future.
540 Delegation Containment
541 ~~~~~~~~~~~~~~~~~~~~~~
543 A delegated sub-hierarchy is contained in the sense that processes
544 can't be moved into or out of the sub-hierarchy by the delegatee.
546 For delegations to a less privileged user, this is achieved by
547 requiring the following conditions for a process with a non-root euid
548 to migrate a target process into a cgroup by writing its PID to the
551 - The writer must have write access to the "cgroup.procs" file.
553 - The writer must have write access to the "cgroup.procs" file of the
554 common ancestor of the source and destination cgroups.
556 The above two constraints ensure that while a delegatee may migrate
557 processes around freely in the delegated sub-hierarchy it can't pull
558 in from or push out to outside the sub-hierarchy.
560 For an example, let's assume cgroups C0 and C1 have been delegated to
561 user U0 who created C00, C01 under C0 and C10 under C1 as follows and
562 all processes under C0 and C1 belong to U0::
564 ~~~~~~~~~~~~~ - C0 - C00
567 ~~~~~~~~~~~~~ - C1 - C10
569 Let's also say U0 wants to write the PID of a process which is
570 currently in C10 into "C00/cgroup.procs". U0 has write access to the
571 file; however, the common ancestor of the source cgroup C10 and the
572 destination cgroup C00 is above the points of delegation and U0 would
573 not have write access to its "cgroup.procs" files and thus the write
574 will be denied with -EACCES.
576 For delegations to namespaces, containment is achieved by requiring
577 that both the source and destination cgroups are reachable from the
578 namespace of the process which is attempting the migration. If either
579 is not reachable, the migration is rejected with -ENOENT.
585 Organize Once and Control
586 ~~~~~~~~~~~~~~~~~~~~~~~~~
588 Migrating a process across cgroups is a relatively expensive operation
589 and stateful resources such as memory are not moved together with the
590 process. This is an explicit design decision as there often exist
591 inherent trade-offs between migration and various hot paths in terms
592 of synchronization cost.
594 As such, migrating processes across cgroups frequently as a means to
595 apply different resource restrictions is discouraged. A workload
596 should be assigned to a cgroup according to the system's logical and
597 resource structure once on start-up. Dynamic adjustments to resource
598 distribution can be made by changing controller configuration through
602 Avoid Name Collisions
603 ~~~~~~~~~~~~~~~~~~~~~
605 Interface files for a cgroup and its children cgroups occupy the same
606 directory and it is possible to create children cgroups which collide
607 with interface files.
609 All cgroup core interface files are prefixed with "cgroup." and each
610 controller's interface files are prefixed with the controller name and
611 a dot. A controller's name is composed of lower case alphabets and
612 '_'s but never begins with an '_' so it can be used as the prefix
613 character for collision avoidance. Also, interface file names won't
614 start or end with terms which are often used in categorizing workloads
615 such as job, service, slice, unit or workload.
617 cgroup doesn't do anything to prevent name collisions and it's the
618 user's responsibility to avoid them.
621 Resource Distribution Models
622 ============================
624 cgroup controllers implement several resource distribution schemes
625 depending on the resource type and expected use cases. This section
626 describes major schemes in use along with their expected behaviors.
632 A parent's resource is distributed by adding up the weights of all
633 active children and giving each the fraction matching the ratio of its
634 weight against the sum. As only children which can make use of the
635 resource at the moment participate in the distribution, this is
636 work-conserving. Due to the dynamic nature, this model is usually
637 used for stateless resources.
639 All weights are in the range [1, 10000] with the default at 100. This
640 allows symmetric multiplicative biases in both directions at fine
641 enough granularity while staying in the intuitive range.
643 As long as the weight is in range, all configuration combinations are
644 valid and there is no reason to reject configuration changes or
647 "cpu.weight" proportionally distributes CPU cycles to active children
648 and is an example of this type.
651 .. _cgroupv2-limits-distributor:
656 A child can only consume up to the configured amount of the resource.
657 Limits can be over-committed - the sum of the limits of children can
658 exceed the amount of resource available to the parent.
660 Limits are in the range [0, max] and defaults to "max", which is noop.
662 As limits can be over-committed, all configuration combinations are
663 valid and there is no reason to reject configuration changes or
666 "io.max" limits the maximum BPS and/or IOPS that a cgroup can consume
667 on an IO device and is an example of this type.
669 .. _cgroupv2-protections-distributor:
674 A cgroup is protected up to the configured amount of the resource
675 as long as the usages of all its ancestors are under their
676 protected levels. Protections can be hard guarantees or best effort
677 soft boundaries. Protections can also be over-committed in which case
678 only up to the amount available to the parent is protected among
681 Protections are in the range [0, max] and defaults to 0, which is
684 As protections can be over-committed, all configuration combinations
685 are valid and there is no reason to reject configuration changes or
688 "memory.low" implements best-effort memory protection and is an
689 example of this type.
695 A cgroup is exclusively allocated a certain amount of a finite
696 resource. Allocations can't be over-committed - the sum of the
697 allocations of children can not exceed the amount of resource
698 available to the parent.
700 Allocations are in the range [0, max] and defaults to 0, which is no
703 As allocations can't be over-committed, some configuration
704 combinations are invalid and should be rejected. Also, if the
705 resource is mandatory for execution of processes, process migrations
708 "cpu.rt.max" hard-allocates realtime slices and is an example of this
718 All interface files should be in one of the following formats whenever
721 New-line separated values
722 (when only one value can be written at once)
728 Space separated values
729 (when read-only or multiple values can be written at once)
741 KEY0 SUB_KEY0=VAL00 SUB_KEY1=VAL01...
742 KEY1 SUB_KEY0=VAL10 SUB_KEY1=VAL11...
745 For a writable file, the format for writing should generally match
746 reading; however, controllers may allow omitting later fields or
747 implement restricted shortcuts for most common use cases.
749 For both flat and nested keyed files, only the values for a single key
750 can be written at a time. For nested keyed files, the sub key pairs
751 may be specified in any order and not all pairs have to be specified.
757 - Settings for a single feature should be contained in a single file.
759 - The root cgroup should be exempt from resource control and thus
760 shouldn't have resource control interface files.
762 - The default time unit is microseconds. If a different unit is ever
763 used, an explicit unit suffix must be present.
765 - A parts-per quantity should use a percentage decimal with at least
766 two digit fractional part - e.g. 13.40.
768 - If a controller implements weight based resource distribution, its
769 interface file should be named "weight" and have the range [1,
770 10000] with 100 as the default. The values are chosen to allow
771 enough and symmetric bias in both directions while keeping it
772 intuitive (the default is 100%).
774 - If a controller implements an absolute resource guarantee and/or
775 limit, the interface files should be named "min" and "max"
776 respectively. If a controller implements best effort resource
777 guarantee and/or limit, the interface files should be named "low"
778 and "high" respectively.
780 In the above four control files, the special token "max" should be
781 used to represent upward infinity for both reading and writing.
783 - If a setting has a configurable default value and keyed specific
784 overrides, the default entry should be keyed with "default" and
785 appear as the first entry in the file.
787 The default value can be updated by writing either "default $VAL" or
790 When writing to update a specific override, "default" can be used as
791 the value to indicate removal of the override. Override entries
792 with "default" as the value must not appear when read.
794 For example, a setting which is keyed by major:minor device numbers
795 with integer values may look like the following::
797 # cat cgroup-example-interface-file
801 The default value can be updated by::
803 # echo 125 > cgroup-example-interface-file
807 # echo "default 125" > cgroup-example-interface-file
809 An override can be set by::
811 # echo "8:16 170" > cgroup-example-interface-file
815 # echo "8:0 default" > cgroup-example-interface-file
816 # cat cgroup-example-interface-file
820 - For events which are not very high frequency, an interface file
821 "events" should be created which lists event key value pairs.
822 Whenever a notifiable event happens, file modified event should be
823 generated on the file.
829 All cgroup core files are prefixed with "cgroup."
832 A read-write single value file which exists on non-root
835 When read, it indicates the current type of the cgroup, which
836 can be one of the following values.
838 - "domain" : A normal valid domain cgroup.
840 - "domain threaded" : A threaded domain cgroup which is
841 serving as the root of a threaded subtree.
843 - "domain invalid" : A cgroup which is in an invalid state.
844 It can't be populated or have controllers enabled. It may
845 be allowed to become a threaded cgroup.
847 - "threaded" : A threaded cgroup which is a member of a
850 A cgroup can be turned into a threaded cgroup by writing
851 "threaded" to this file.
854 A read-write new-line separated values file which exists on
857 When read, it lists the PIDs of all processes which belong to
858 the cgroup one-per-line. The PIDs are not ordered and the
859 same PID may show up more than once if the process got moved
860 to another cgroup and then back or the PID got recycled while
863 A PID can be written to migrate the process associated with
864 the PID to the cgroup. The writer should match all of the
865 following conditions.
867 - It must have write access to the "cgroup.procs" file.
869 - It must have write access to the "cgroup.procs" file of the
870 common ancestor of the source and destination cgroups.
872 When delegating a sub-hierarchy, write access to this file
873 should be granted along with the containing directory.
875 In a threaded cgroup, reading this file fails with EOPNOTSUPP
876 as all the processes belong to the thread root. Writing is
877 supported and moves every thread of the process to the cgroup.
880 A read-write new-line separated values file which exists on
883 When read, it lists the TIDs of all threads which belong to
884 the cgroup one-per-line. The TIDs are not ordered and the
885 same TID may show up more than once if the thread got moved to
886 another cgroup and then back or the TID got recycled while
889 A TID can be written to migrate the thread associated with the
890 TID to the cgroup. The writer should match all of the
891 following conditions.
893 - It must have write access to the "cgroup.threads" file.
895 - The cgroup that the thread is currently in must be in the
896 same resource domain as the destination cgroup.
898 - It must have write access to the "cgroup.procs" file of the
899 common ancestor of the source and destination cgroups.
901 When delegating a sub-hierarchy, write access to this file
902 should be granted along with the containing directory.
905 A read-only space separated values file which exists on all
908 It shows space separated list of all controllers available to
909 the cgroup. The controllers are not ordered.
911 cgroup.subtree_control
912 A read-write space separated values file which exists on all
913 cgroups. Starts out empty.
915 When read, it shows space separated list of the controllers
916 which are enabled to control resource distribution from the
917 cgroup to its children.
919 Space separated list of controllers prefixed with '+' or '-'
920 can be written to enable or disable controllers. A controller
921 name prefixed with '+' enables the controller and '-'
922 disables. If a controller appears more than once on the list,
923 the last one is effective. When multiple enable and disable
924 operations are specified, either all succeed or all fail.
927 A read-only flat-keyed file which exists on non-root cgroups.
928 The following entries are defined. Unless specified
929 otherwise, a value change in this file generates a file
933 1 if the cgroup or its descendants contains any live
934 processes; otherwise, 0.
936 1 if the cgroup is frozen; otherwise, 0.
938 cgroup.max.descendants
939 A read-write single value files. The default is "max".
941 Maximum allowed number of descent cgroups.
942 If the actual number of descendants is equal or larger,
943 an attempt to create a new cgroup in the hierarchy will fail.
946 A read-write single value files. The default is "max".
948 Maximum allowed descent depth below the current cgroup.
949 If the actual descent depth is equal or larger,
950 an attempt to create a new child cgroup will fail.
953 A read-only flat-keyed file with the following entries:
956 Total number of visible descendant cgroups.
959 Total number of dying descendant cgroups. A cgroup becomes
960 dying after being deleted by a user. The cgroup will remain
961 in dying state for some time undefined time (which can depend
962 on system load) before being completely destroyed.
964 A process can't enter a dying cgroup under any circumstances,
965 a dying cgroup can't revive.
967 A dying cgroup can consume system resources not exceeding
968 limits, which were active at the moment of cgroup deletion.
971 A read-write single value file which exists on non-root cgroups.
972 Allowed values are "0" and "1". The default is "0".
974 Writing "1" to the file causes freezing of the cgroup and all
975 descendant cgroups. This means that all belonging processes will
976 be stopped and will not run until the cgroup will be explicitly
977 unfrozen. Freezing of the cgroup may take some time; when this action
978 is completed, the "frozen" value in the cgroup.events control file
979 will be updated to "1" and the corresponding notification will be
982 A cgroup can be frozen either by its own settings, or by settings
983 of any ancestor cgroups. If any of ancestor cgroups is frozen, the
984 cgroup will remain frozen.
986 Processes in the frozen cgroup can be killed by a fatal signal.
987 They also can enter and leave a frozen cgroup: either by an explicit
988 move by a user, or if freezing of the cgroup races with fork().
989 If a process is moved to a frozen cgroup, it stops. If a process is
990 moved out of a frozen cgroup, it becomes running.
992 Frozen status of a cgroup doesn't affect any cgroup tree operations:
993 it's possible to delete a frozen (and empty) cgroup, as well as
994 create new sub-cgroups.
997 A write-only single value file which exists in non-root cgroups.
998 The only allowed value is "1".
1000 Writing "1" to the file causes the cgroup and all descendant cgroups to
1001 be killed. This means that all processes located in the affected cgroup
1002 tree will be killed via SIGKILL.
1004 Killing a cgroup tree will deal with concurrent forks appropriately and
1005 is protected against migrations.
1007 In a threaded cgroup, writing this file fails with EOPNOTSUPP as
1008 killing cgroups is a process directed operation, i.e. it affects
1009 the whole thread-group.
1012 A read-write single value file that allowed values are "0" and "1".
1015 Writing "0" to the file will disable the cgroup PSI accounting.
1016 Writing "1" to the file will re-enable the cgroup PSI accounting.
1018 This control attribute is not hierarchical, so disable or enable PSI
1019 accounting in a cgroup does not affect PSI accounting in descendants
1020 and doesn't need pass enablement via ancestors from root.
1022 The reason this control attribute exists is that PSI accounts stalls for
1023 each cgroup separately and aggregates it at each level of the hierarchy.
1024 This may cause non-negligible overhead for some workloads when under
1025 deep level of the hierarchy, in which case this control attribute can
1026 be used to disable PSI accounting in the non-leaf cgroups.
1029 A read-write nested-keyed file.
1031 Shows pressure stall information for IRQ/SOFTIRQ. See
1032 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1042 The "cpu" controllers regulates distribution of CPU cycles. This
1043 controller implements weight and absolute bandwidth limit models for
1044 normal scheduling policy and absolute bandwidth allocation model for
1045 realtime scheduling policy.
1047 In all the above models, cycles distribution is defined only on a temporal
1048 base and it does not account for the frequency at which tasks are executed.
1049 The (optional) utilization clamping support allows to hint the schedutil
1050 cpufreq governor about the minimum desired frequency which should always be
1051 provided by a CPU, as well as the maximum desired frequency, which should not
1052 be exceeded by a CPU.
1054 WARNING: cgroup2 doesn't yet support control of realtime processes and
1055 the cpu controller can only be enabled when all RT processes are in
1056 the root cgroup. Be aware that system management software may already
1057 have placed RT processes into nonroot cgroups during the system boot
1058 process, and these processes may need to be moved to the root cgroup
1059 before the cpu controller can be enabled.
1065 All time durations are in microseconds.
1068 A read-only flat-keyed file.
1069 This file exists whether the controller is enabled or not.
1071 It always reports the following three stats:
1077 and the following five when the controller is enabled:
1086 A read-write single value file which exists on non-root
1087 cgroups. The default is "100".
1089 The weight in the range [1, 10000].
1092 A read-write single value file which exists on non-root
1093 cgroups. The default is "0".
1095 The nice value is in the range [-20, 19].
1097 This interface file is an alternative interface for
1098 "cpu.weight" and allows reading and setting weight using the
1099 same values used by nice(2). Because the range is smaller and
1100 granularity is coarser for the nice values, the read value is
1101 the closest approximation of the current weight.
1104 A read-write two value file which exists on non-root cgroups.
1105 The default is "max 100000".
1107 The maximum bandwidth limit. It's in the following format::
1111 which indicates that the group may consume up to $MAX in each
1112 $PERIOD duration. "max" for $MAX indicates no limit. If only
1113 one number is written, $MAX is updated.
1116 A read-write single value file which exists on non-root
1117 cgroups. The default is "0".
1119 The burst in the range [0, $MAX].
1122 A read-write nested-keyed file.
1124 Shows pressure stall information for CPU. See
1125 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1128 A read-write single value file which exists on non-root cgroups.
1129 The default is "0", i.e. no utilization boosting.
1131 The requested minimum utilization (protection) as a percentage
1132 rational number, e.g. 12.34 for 12.34%.
1134 This interface allows reading and setting minimum utilization clamp
1135 values similar to the sched_setattr(2). This minimum utilization
1136 value is used to clamp the task specific minimum utilization clamp.
1138 The requested minimum utilization (protection) is always capped by
1139 the current value for the maximum utilization (limit), i.e.
1143 A read-write single value file which exists on non-root cgroups.
1144 The default is "max". i.e. no utilization capping
1146 The requested maximum utilization (limit) as a percentage rational
1147 number, e.g. 98.76 for 98.76%.
1149 This interface allows reading and setting maximum utilization clamp
1150 values similar to the sched_setattr(2). This maximum utilization
1151 value is used to clamp the task specific maximum utilization clamp.
1158 The "memory" controller regulates distribution of memory. Memory is
1159 stateful and implements both limit and protection models. Due to the
1160 intertwining between memory usage and reclaim pressure and the
1161 stateful nature of memory, the distribution model is relatively
1164 While not completely water-tight, all major memory usages by a given
1165 cgroup are tracked so that the total memory consumption can be
1166 accounted and controlled to a reasonable extent. Currently, the
1167 following types of memory usages are tracked.
1169 - Userland memory - page cache and anonymous memory.
1171 - Kernel data structures such as dentries and inodes.
1173 - TCP socket buffers.
1175 The above list may expand in the future for better coverage.
1178 Memory Interface Files
1179 ~~~~~~~~~~~~~~~~~~~~~~
1181 All memory amounts are in bytes. If a value which is not aligned to
1182 PAGE_SIZE is written, the value may be rounded up to the closest
1183 PAGE_SIZE multiple when read back.
1186 A read-only single value file which exists on non-root
1189 The total amount of memory currently being used by the cgroup
1190 and its descendants.
1193 A read-write single value file which exists on non-root
1194 cgroups. The default is "0".
1196 Hard memory protection. If the memory usage of a cgroup
1197 is within its effective min boundary, the cgroup's memory
1198 won't be reclaimed under any conditions. If there is no
1199 unprotected reclaimable memory available, OOM killer
1200 is invoked. Above the effective min boundary (or
1201 effective low boundary if it is higher), pages are reclaimed
1202 proportionally to the overage, reducing reclaim pressure for
1205 Effective min boundary is limited by memory.min values of
1206 all ancestor cgroups. If there is memory.min overcommitment
1207 (child cgroup or cgroups are requiring more protected memory
1208 than parent will allow), then each child cgroup will get
1209 the part of parent's protection proportional to its
1210 actual memory usage below memory.min.
1212 Putting more memory than generally available under this
1213 protection is discouraged and may lead to constant OOMs.
1215 If a memory cgroup is not populated with processes,
1216 its memory.min is ignored.
1219 A read-write single value file which exists on non-root
1220 cgroups. The default is "0".
1222 Best-effort memory protection. If the memory usage of a
1223 cgroup is within its effective low boundary, the cgroup's
1224 memory won't be reclaimed unless there is no reclaimable
1225 memory available in unprotected cgroups.
1226 Above the effective low boundary (or
1227 effective min boundary if it is higher), pages are reclaimed
1228 proportionally to the overage, reducing reclaim pressure for
1231 Effective low boundary is limited by memory.low values of
1232 all ancestor cgroups. If there is memory.low overcommitment
1233 (child cgroup or cgroups are requiring more protected memory
1234 than parent will allow), then each child cgroup will get
1235 the part of parent's protection proportional to its
1236 actual memory usage below memory.low.
1238 Putting more memory than generally available under this
1239 protection is discouraged.
1242 A read-write single value file which exists on non-root
1243 cgroups. The default is "max".
1245 Memory usage throttle limit. If a cgroup's usage goes
1246 over the high boundary, the processes of the cgroup are
1247 throttled and put under heavy reclaim pressure.
1249 Going over the high limit never invokes the OOM killer and
1250 under extreme conditions the limit may be breached. The high
1251 limit should be used in scenarios where an external process
1252 monitors the limited cgroup to alleviate heavy reclaim
1256 A read-write single value file which exists on non-root
1257 cgroups. The default is "max".
1259 Memory usage hard limit. This is the main mechanism to limit
1260 memory usage of a cgroup. If a cgroup's memory usage reaches
1261 this limit and can't be reduced, the OOM killer is invoked in
1262 the cgroup. Under certain circumstances, the usage may go
1263 over the limit temporarily.
1265 In default configuration regular 0-order allocations always
1266 succeed unless OOM killer chooses current task as a victim.
1268 Some kinds of allocations don't invoke the OOM killer.
1269 Caller could retry them differently, return into userspace
1270 as -ENOMEM or silently ignore in cases like disk readahead.
1273 A write-only nested-keyed file which exists for all cgroups.
1275 This is a simple interface to trigger memory reclaim in the
1278 This file accepts a single key, the number of bytes to reclaim.
1279 No nested keys are currently supported.
1283 echo "1G" > memory.reclaim
1285 The interface can be later extended with nested keys to
1286 configure the reclaim behavior. For example, specify the
1287 type of memory to reclaim from (anon, file, ..).
1289 Please note that the kernel can over or under reclaim from
1290 the target cgroup. If less bytes are reclaimed than the
1291 specified amount, -EAGAIN is returned.
1293 Please note that the proactive reclaim (triggered by this
1294 interface) is not meant to indicate memory pressure on the
1295 memory cgroup. Therefore socket memory balancing triggered by
1296 the memory reclaim normally is not exercised in this case.
1297 This means that the networking layer will not adapt based on
1298 reclaim induced by memory.reclaim.
1301 A read-only single value file which exists on non-root
1304 The max memory usage recorded for the cgroup and its
1305 descendants since the creation of the cgroup.
1308 A read-write single value file which exists on non-root
1309 cgroups. The default value is "0".
1311 Determines whether the cgroup should be treated as
1312 an indivisible workload by the OOM killer. If set,
1313 all tasks belonging to the cgroup or to its descendants
1314 (if the memory cgroup is not a leaf cgroup) are killed
1315 together or not at all. This can be used to avoid
1316 partial kills to guarantee workload integrity.
1318 Tasks with the OOM protection (oom_score_adj set to -1000)
1319 are treated as an exception and are never killed.
1321 If the OOM killer is invoked in a cgroup, it's not going
1322 to kill any tasks outside of this cgroup, regardless
1323 memory.oom.group values of ancestor cgroups.
1326 A read-only flat-keyed file which exists on non-root cgroups.
1327 The following entries are defined. Unless specified
1328 otherwise, a value change in this file generates a file
1331 Note that all fields in this file are hierarchical and the
1332 file modified event can be generated due to an event down the
1333 hierarchy. For the local events at the cgroup level see
1334 memory.events.local.
1337 The number of times the cgroup is reclaimed due to
1338 high memory pressure even though its usage is under
1339 the low boundary. This usually indicates that the low
1340 boundary is over-committed.
1343 The number of times processes of the cgroup are
1344 throttled and routed to perform direct memory reclaim
1345 because the high memory boundary was exceeded. For a
1346 cgroup whose memory usage is capped by the high limit
1347 rather than global memory pressure, this event's
1348 occurrences are expected.
1351 The number of times the cgroup's memory usage was
1352 about to go over the max boundary. If direct reclaim
1353 fails to bring it down, the cgroup goes to OOM state.
1356 The number of time the cgroup's memory usage was
1357 reached the limit and allocation was about to fail.
1359 This event is not raised if the OOM killer is not
1360 considered as an option, e.g. for failed high-order
1361 allocations or if caller asked to not retry attempts.
1364 The number of processes belonging to this cgroup
1365 killed by any kind of OOM killer.
1368 The number of times a group OOM has occurred.
1371 Similar to memory.events but the fields in the file are local
1372 to the cgroup i.e. not hierarchical. The file modified event
1373 generated on this file reflects only the local events.
1376 A read-only flat-keyed file which exists on non-root cgroups.
1378 This breaks down the cgroup's memory footprint into different
1379 types of memory, type-specific details, and other information
1380 on the state and past events of the memory management system.
1382 All memory amounts are in bytes.
1384 The entries are ordered to be human readable, and new entries
1385 can show up in the middle. Don't rely on items remaining in a
1386 fixed position; use the keys to look up specific values!
1388 If the entry has no per-node counter (or not show in the
1389 memory.numa_stat). We use 'npn' (non-per-node) as the tag
1390 to indicate that it will not show in the memory.numa_stat.
1393 Amount of memory used in anonymous mappings such as
1394 brk(), sbrk(), and mmap(MAP_ANONYMOUS)
1397 Amount of memory used to cache filesystem data,
1398 including tmpfs and shared memory.
1401 Amount of total kernel memory, including
1402 (kernel_stack, pagetables, percpu, vmalloc, slab) in
1403 addition to other kernel memory use cases.
1406 Amount of memory allocated to kernel stacks.
1409 Amount of memory allocated for page tables.
1412 Amount of memory allocated for secondary page tables,
1413 this currently includes KVM mmu allocations on x86
1417 Amount of memory used for storing per-cpu kernel
1421 Amount of memory used in network transmission buffers
1424 Amount of memory used for vmap backed memory.
1427 Amount of cached filesystem data that is swap-backed,
1428 such as tmpfs, shm segments, shared anonymous mmap()s
1431 Amount of memory consumed by the zswap compression backend.
1434 Amount of application memory swapped out to zswap.
1437 Amount of cached filesystem data mapped with mmap()
1440 Amount of cached filesystem data that was modified but
1441 not yet written back to disk
1444 Amount of cached filesystem data that was modified and
1445 is currently being written back to disk
1448 Amount of swap cached in memory. The swapcache is accounted
1449 against both memory and swap usage.
1452 Amount of memory used in anonymous mappings backed by
1453 transparent hugepages
1456 Amount of cached filesystem data backed by transparent
1460 Amount of shm, tmpfs, shared anonymous mmap()s backed by
1461 transparent hugepages
1463 inactive_anon, active_anon, inactive_file, active_file, unevictable
1464 Amount of memory, swap-backed and filesystem-backed,
1465 on the internal memory management lists used by the
1466 page reclaim algorithm.
1468 As these represent internal list state (eg. shmem pages are on anon
1469 memory management lists), inactive_foo + active_foo may not be equal to
1470 the value for the foo counter, since the foo counter is type-based, not
1474 Part of "slab" that might be reclaimed, such as
1475 dentries and inodes.
1478 Part of "slab" that cannot be reclaimed on memory
1482 Amount of memory used for storing in-kernel data
1485 workingset_refault_anon
1486 Number of refaults of previously evicted anonymous pages.
1488 workingset_refault_file
1489 Number of refaults of previously evicted file pages.
1491 workingset_activate_anon
1492 Number of refaulted anonymous pages that were immediately
1495 workingset_activate_file
1496 Number of refaulted file pages that were immediately activated.
1498 workingset_restore_anon
1499 Number of restored anonymous pages which have been detected as
1500 an active workingset before they got reclaimed.
1502 workingset_restore_file
1503 Number of restored file pages which have been detected as an
1504 active workingset before they got reclaimed.
1506 workingset_nodereclaim
1507 Number of times a shadow node has been reclaimed
1510 Amount of scanned pages (in an inactive LRU list)
1513 Amount of reclaimed pages
1516 Amount of scanned pages by kswapd (in an inactive LRU list)
1519 Amount of scanned pages directly (in an inactive LRU list)
1521 pgscan_khugepaged (npn)
1522 Amount of scanned pages by khugepaged (in an inactive LRU list)
1524 pgsteal_kswapd (npn)
1525 Amount of reclaimed pages by kswapd
1527 pgsteal_direct (npn)
1528 Amount of reclaimed pages directly
1530 pgsteal_khugepaged (npn)
1531 Amount of reclaimed pages by khugepaged
1534 Total number of page faults incurred
1537 Number of major page faults incurred
1540 Amount of scanned pages (in an active LRU list)
1543 Amount of pages moved to the active LRU list
1546 Amount of pages moved to the inactive LRU list
1549 Amount of pages postponed to be freed under memory pressure
1552 Amount of reclaimed lazyfree pages
1554 thp_fault_alloc (npn)
1555 Number of transparent hugepages which were allocated to satisfy
1556 a page fault. This counter is not present when CONFIG_TRANSPARENT_HUGEPAGE
1559 thp_collapse_alloc (npn)
1560 Number of transparent hugepages which were allocated to allow
1561 collapsing an existing range of pages. This counter is not
1562 present when CONFIG_TRANSPARENT_HUGEPAGE is not set.
1565 Number of transparent hugepages which are swapout in one piece
1568 thp_swpout_fallback (npn)
1569 Number of transparent hugepages which were split before swapout.
1570 Usually because failed to allocate some continuous swap space
1574 A read-only nested-keyed file which exists on non-root cgroups.
1576 This breaks down the cgroup's memory footprint into different
1577 types of memory, type-specific details, and other information
1578 per node on the state of the memory management system.
1580 This is useful for providing visibility into the NUMA locality
1581 information within an memcg since the pages are allowed to be
1582 allocated from any physical node. One of the use case is evaluating
1583 application performance by combining this information with the
1584 application's CPU allocation.
1586 All memory amounts are in bytes.
1588 The output format of memory.numa_stat is::
1590 type N0=<bytes in node 0> N1=<bytes in node 1> ...
1592 The entries are ordered to be human readable, and new entries
1593 can show up in the middle. Don't rely on items remaining in a
1594 fixed position; use the keys to look up specific values!
1596 The entries can refer to the memory.stat.
1599 A read-only single value file which exists on non-root
1602 The total amount of swap currently being used by the cgroup
1603 and its descendants.
1606 A read-write single value file which exists on non-root
1607 cgroups. The default is "max".
1609 Swap usage throttle limit. If a cgroup's swap usage exceeds
1610 this limit, all its further allocations will be throttled to
1611 allow userspace to implement custom out-of-memory procedures.
1613 This limit marks a point of no return for the cgroup. It is NOT
1614 designed to manage the amount of swapping a workload does
1615 during regular operation. Compare to memory.swap.max, which
1616 prohibits swapping past a set amount, but lets the cgroup
1617 continue unimpeded as long as other memory can be reclaimed.
1619 Healthy workloads are not expected to reach this limit.
1622 A read-only single value file which exists on non-root
1625 The max swap usage recorded for the cgroup and its
1626 descendants since the creation of the cgroup.
1629 A read-write single value file which exists on non-root
1630 cgroups. The default is "max".
1632 Swap usage hard limit. If a cgroup's swap usage reaches this
1633 limit, anonymous memory of the cgroup will not be swapped out.
1636 A read-only flat-keyed file which exists on non-root cgroups.
1637 The following entries are defined. Unless specified
1638 otherwise, a value change in this file generates a file
1642 The number of times the cgroup's swap usage was over
1646 The number of times the cgroup's swap usage was about
1647 to go over the max boundary and swap allocation
1651 The number of times swap allocation failed either
1652 because of running out of swap system-wide or max
1655 When reduced under the current usage, the existing swap
1656 entries are reclaimed gradually and the swap usage may stay
1657 higher than the limit for an extended period of time. This
1658 reduces the impact on the workload and memory management.
1660 memory.zswap.current
1661 A read-only single value file which exists on non-root
1664 The total amount of memory consumed by the zswap compression
1668 A read-write single value file which exists on non-root
1669 cgroups. The default is "max".
1671 Zswap usage hard limit. If a cgroup's zswap pool reaches this
1672 limit, it will refuse to take any more stores before existing
1673 entries fault back in or are written out to disk.
1676 A read-only nested-keyed file.
1678 Shows pressure stall information for memory. See
1679 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1685 "memory.high" is the main mechanism to control memory usage.
1686 Over-committing on high limit (sum of high limits > available memory)
1687 and letting global memory pressure to distribute memory according to
1688 usage is a viable strategy.
1690 Because breach of the high limit doesn't trigger the OOM killer but
1691 throttles the offending cgroup, a management agent has ample
1692 opportunities to monitor and take appropriate actions such as granting
1693 more memory or terminating the workload.
1695 Determining whether a cgroup has enough memory is not trivial as
1696 memory usage doesn't indicate whether the workload can benefit from
1697 more memory. For example, a workload which writes data received from
1698 network to a file can use all available memory but can also operate as
1699 performant with a small amount of memory. A measure of memory
1700 pressure - how much the workload is being impacted due to lack of
1701 memory - is necessary to determine whether a workload needs more
1702 memory; unfortunately, memory pressure monitoring mechanism isn't
1709 A memory area is charged to the cgroup which instantiated it and stays
1710 charged to the cgroup until the area is released. Migrating a process
1711 to a different cgroup doesn't move the memory usages that it
1712 instantiated while in the previous cgroup to the new cgroup.
1714 A memory area may be used by processes belonging to different cgroups.
1715 To which cgroup the area will be charged is in-deterministic; however,
1716 over time, the memory area is likely to end up in a cgroup which has
1717 enough memory allowance to avoid high reclaim pressure.
1719 If a cgroup sweeps a considerable amount of memory which is expected
1720 to be accessed repeatedly by other cgroups, it may make sense to use
1721 POSIX_FADV_DONTNEED to relinquish the ownership of memory areas
1722 belonging to the affected files to ensure correct memory ownership.
1728 The "io" controller regulates the distribution of IO resources. This
1729 controller implements both weight based and absolute bandwidth or IOPS
1730 limit distribution; however, weight based distribution is available
1731 only if cfq-iosched is in use and neither scheme is available for
1739 A read-only nested-keyed file.
1741 Lines are keyed by $MAJ:$MIN device numbers and not ordered.
1742 The following nested keys are defined.
1744 ====== =====================
1746 wbytes Bytes written
1747 rios Number of read IOs
1748 wios Number of write IOs
1749 dbytes Bytes discarded
1750 dios Number of discard IOs
1751 ====== =====================
1753 An example read output follows::
1755 8:16 rbytes=1459200 wbytes=314773504 rios=192 wios=353 dbytes=0 dios=0
1756 8:0 rbytes=90430464 wbytes=299008000 rios=8950 wios=1252 dbytes=50331648 dios=3021
1759 A read-write nested-keyed file which exists only on the root
1762 This file configures the Quality of Service of the IO cost
1763 model based controller (CONFIG_BLK_CGROUP_IOCOST) which
1764 currently implements "io.weight" proportional control. Lines
1765 are keyed by $MAJ:$MIN device numbers and not ordered. The
1766 line for a given device is populated on the first write for
1767 the device on "io.cost.qos" or "io.cost.model". The following
1768 nested keys are defined.
1770 ====== =====================================
1771 enable Weight-based control enable
1772 ctrl "auto" or "user"
1773 rpct Read latency percentile [0, 100]
1774 rlat Read latency threshold
1775 wpct Write latency percentile [0, 100]
1776 wlat Write latency threshold
1777 min Minimum scaling percentage [1, 10000]
1778 max Maximum scaling percentage [1, 10000]
1779 ====== =====================================
1781 The controller is disabled by default and can be enabled by
1782 setting "enable" to 1. "rpct" and "wpct" parameters default
1783 to zero and the controller uses internal device saturation
1784 state to adjust the overall IO rate between "min" and "max".
1786 When a better control quality is needed, latency QoS
1787 parameters can be configured. For example::
1789 8:16 enable=1 ctrl=auto rpct=95.00 rlat=75000 wpct=95.00 wlat=150000 min=50.00 max=150.0
1791 shows that on sdb, the controller is enabled, will consider
1792 the device saturated if the 95th percentile of read completion
1793 latencies is above 75ms or write 150ms, and adjust the overall
1794 IO issue rate between 50% and 150% accordingly.
1796 The lower the saturation point, the better the latency QoS at
1797 the cost of aggregate bandwidth. The narrower the allowed
1798 adjustment range between "min" and "max", the more conformant
1799 to the cost model the IO behavior. Note that the IO issue
1800 base rate may be far off from 100% and setting "min" and "max"
1801 blindly can lead to a significant loss of device capacity or
1802 control quality. "min" and "max" are useful for regulating
1803 devices which show wide temporary behavior changes - e.g. a
1804 ssd which accepts writes at the line speed for a while and
1805 then completely stalls for multiple seconds.
1807 When "ctrl" is "auto", the parameters are controlled by the
1808 kernel and may change automatically. Setting "ctrl" to "user"
1809 or setting any of the percentile and latency parameters puts
1810 it into "user" mode and disables the automatic changes. The
1811 automatic mode can be restored by setting "ctrl" to "auto".
1814 A read-write nested-keyed file which exists only on the root
1817 This file configures the cost model of the IO cost model based
1818 controller (CONFIG_BLK_CGROUP_IOCOST) which currently
1819 implements "io.weight" proportional control. Lines are keyed
1820 by $MAJ:$MIN device numbers and not ordered. The line for a
1821 given device is populated on the first write for the device on
1822 "io.cost.qos" or "io.cost.model". The following nested keys
1825 ===== ================================
1826 ctrl "auto" or "user"
1827 model The cost model in use - "linear"
1828 ===== ================================
1830 When "ctrl" is "auto", the kernel may change all parameters
1831 dynamically. When "ctrl" is set to "user" or any other
1832 parameters are written to, "ctrl" become "user" and the
1833 automatic changes are disabled.
1835 When "model" is "linear", the following model parameters are
1838 ============= ========================================
1839 [r|w]bps The maximum sequential IO throughput
1840 [r|w]seqiops The maximum 4k sequential IOs per second
1841 [r|w]randiops The maximum 4k random IOs per second
1842 ============= ========================================
1844 From the above, the builtin linear model determines the base
1845 costs of a sequential and random IO and the cost coefficient
1846 for the IO size. While simple, this model can cover most
1847 common device classes acceptably.
1849 The IO cost model isn't expected to be accurate in absolute
1850 sense and is scaled to the device behavior dynamically.
1852 If needed, tools/cgroup/iocost_coef_gen.py can be used to
1853 generate device-specific coefficients.
1856 A read-write flat-keyed file which exists on non-root cgroups.
1857 The default is "default 100".
1859 The first line is the default weight applied to devices
1860 without specific override. The rest are overrides keyed by
1861 $MAJ:$MIN device numbers and not ordered. The weights are in
1862 the range [1, 10000] and specifies the relative amount IO time
1863 the cgroup can use in relation to its siblings.
1865 The default weight can be updated by writing either "default
1866 $WEIGHT" or simply "$WEIGHT". Overrides can be set by writing
1867 "$MAJ:$MIN $WEIGHT" and unset by writing "$MAJ:$MIN default".
1869 An example read output follows::
1876 A read-write nested-keyed file which exists on non-root
1879 BPS and IOPS based IO limit. Lines are keyed by $MAJ:$MIN
1880 device numbers and not ordered. The following nested keys are
1883 ===== ==================================
1884 rbps Max read bytes per second
1885 wbps Max write bytes per second
1886 riops Max read IO operations per second
1887 wiops Max write IO operations per second
1888 ===== ==================================
1890 When writing, any number of nested key-value pairs can be
1891 specified in any order. "max" can be specified as the value
1892 to remove a specific limit. If the same key is specified
1893 multiple times, the outcome is undefined.
1895 BPS and IOPS are measured in each IO direction and IOs are
1896 delayed if limit is reached. Temporary bursts are allowed.
1898 Setting read limit at 2M BPS and write at 120 IOPS for 8:16::
1900 echo "8:16 rbps=2097152 wiops=120" > io.max
1902 Reading returns the following::
1904 8:16 rbps=2097152 wbps=max riops=max wiops=120
1906 Write IOPS limit can be removed by writing the following::
1908 echo "8:16 wiops=max" > io.max
1910 Reading now returns the following::
1912 8:16 rbps=2097152 wbps=max riops=max wiops=max
1915 A read-only nested-keyed file.
1917 Shows pressure stall information for IO. See
1918 :ref:`Documentation/accounting/psi.rst <psi>` for details.
1924 Page cache is dirtied through buffered writes and shared mmaps and
1925 written asynchronously to the backing filesystem by the writeback
1926 mechanism. Writeback sits between the memory and IO domains and
1927 regulates the proportion of dirty memory by balancing dirtying and
1930 The io controller, in conjunction with the memory controller,
1931 implements control of page cache writeback IOs. The memory controller
1932 defines the memory domain that dirty memory ratio is calculated and
1933 maintained for and the io controller defines the io domain which
1934 writes out dirty pages for the memory domain. Both system-wide and
1935 per-cgroup dirty memory states are examined and the more restrictive
1936 of the two is enforced.
1938 cgroup writeback requires explicit support from the underlying
1939 filesystem. Currently, cgroup writeback is implemented on ext2, ext4,
1940 btrfs, f2fs, and xfs. On other filesystems, all writeback IOs are
1941 attributed to the root cgroup.
1943 There are inherent differences in memory and writeback management
1944 which affects how cgroup ownership is tracked. Memory is tracked per
1945 page while writeback per inode. For the purpose of writeback, an
1946 inode is assigned to a cgroup and all IO requests to write dirty pages
1947 from the inode are attributed to that cgroup.
1949 As cgroup ownership for memory is tracked per page, there can be pages
1950 which are associated with different cgroups than the one the inode is
1951 associated with. These are called foreign pages. The writeback
1952 constantly keeps track of foreign pages and, if a particular foreign
1953 cgroup becomes the majority over a certain period of time, switches
1954 the ownership of the inode to that cgroup.
1956 While this model is enough for most use cases where a given inode is
1957 mostly dirtied by a single cgroup even when the main writing cgroup
1958 changes over time, use cases where multiple cgroups write to a single
1959 inode simultaneously are not supported well. In such circumstances, a
1960 significant portion of IOs are likely to be attributed incorrectly.
1961 As memory controller assigns page ownership on the first use and
1962 doesn't update it until the page is released, even if writeback
1963 strictly follows page ownership, multiple cgroups dirtying overlapping
1964 areas wouldn't work as expected. It's recommended to avoid such usage
1967 The sysctl knobs which affect writeback behavior are applied to cgroup
1968 writeback as follows.
1970 vm.dirty_background_ratio, vm.dirty_ratio
1971 These ratios apply the same to cgroup writeback with the
1972 amount of available memory capped by limits imposed by the
1973 memory controller and system-wide clean memory.
1975 vm.dirty_background_bytes, vm.dirty_bytes
1976 For cgroup writeback, this is calculated into ratio against
1977 total available memory and applied the same way as
1978 vm.dirty[_background]_ratio.
1984 This is a cgroup v2 controller for IO workload protection. You provide a group
1985 with a latency target, and if the average latency exceeds that target the
1986 controller will throttle any peers that have a lower latency target than the
1989 The limits are only applied at the peer level in the hierarchy. This means that
1990 in the diagram below, only groups A, B, and C will influence each other, and
1991 groups D and F will influence each other. Group G will influence nobody::
2000 So the ideal way to configure this is to set io.latency in groups A, B, and C.
2001 Generally you do not want to set a value lower than the latency your device
2002 supports. Experiment to find the value that works best for your workload.
2003 Start at higher than the expected latency for your device and watch the
2004 avg_lat value in io.stat for your workload group to get an idea of the
2005 latency you see during normal operation. Use the avg_lat value as a basis for
2006 your real setting, setting at 10-15% higher than the value in io.stat.
2008 How IO Latency Throttling Works
2009 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2011 io.latency is work conserving; so as long as everybody is meeting their latency
2012 target the controller doesn't do anything. Once a group starts missing its
2013 target it begins throttling any peer group that has a higher target than itself.
2014 This throttling takes 2 forms:
2016 - Queue depth throttling. This is the number of outstanding IO's a group is
2017 allowed to have. We will clamp down relatively quickly, starting at no limit
2018 and going all the way down to 1 IO at a time.
2020 - Artificial delay induction. There are certain types of IO that cannot be
2021 throttled without possibly adversely affecting higher priority groups. This
2022 includes swapping and metadata IO. These types of IO are allowed to occur
2023 normally, however they are "charged" to the originating group. If the
2024 originating group is being throttled you will see the use_delay and delay
2025 fields in io.stat increase. The delay value is how many microseconds that are
2026 being added to any process that runs in this group. Because this number can
2027 grow quite large if there is a lot of swapping or metadata IO occurring we
2028 limit the individual delay events to 1 second at a time.
2030 Once the victimized group starts meeting its latency target again it will start
2031 unthrottling any peer groups that were throttled previously. If the victimized
2032 group simply stops doing IO the global counter will unthrottle appropriately.
2034 IO Latency Interface Files
2035 ~~~~~~~~~~~~~~~~~~~~~~~~~~
2038 This takes a similar format as the other controllers.
2040 "MAJOR:MINOR target=<target time in microseconds>"
2043 If the controller is enabled you will see extra stats in io.stat in
2044 addition to the normal ones.
2047 This is the current queue depth for the group.
2050 This is an exponential moving average with a decay rate of 1/exp
2051 bound by the sampling interval. The decay rate interval can be
2052 calculated by multiplying the win value in io.stat by the
2053 corresponding number of samples based on the win value.
2056 The sampling window size in milliseconds. This is the minimum
2057 duration of time between evaluation events. Windows only elapse
2058 with IO activity. Idle periods extend the most recent window.
2063 A single attribute controls the behavior of the I/O priority cgroup policy,
2064 namely the blkio.prio.class attribute. The following values are accepted for
2068 Do not modify the I/O priority class.
2071 For requests that have a non-RT I/O priority class, change it into RT.
2072 Also change the priority level of these requests to 4. Do not modify
2073 the I/O priority of requests that have priority class RT.
2076 For requests that do not have an I/O priority class or that have I/O
2077 priority class RT, change it into BE. Also change the priority level
2078 of these requests to 0. Do not modify the I/O priority class of
2079 requests that have priority class IDLE.
2082 Change the I/O priority class of all requests into IDLE, the lowest
2086 Deprecated. Just an alias for promote-to-rt.
2088 The following numerical values are associated with the I/O priority policies:
2090 +----------------+---+
2092 +----------------+---+
2094 +----------------+---+
2096 +----------------+---+
2098 The numerical value that corresponds to each I/O priority class is as follows:
2100 +-------------------------------+---+
2101 | IOPRIO_CLASS_NONE | 0 |
2102 +-------------------------------+---+
2103 | IOPRIO_CLASS_RT (real-time) | 1 |
2104 +-------------------------------+---+
2105 | IOPRIO_CLASS_BE (best effort) | 2 |
2106 +-------------------------------+---+
2107 | IOPRIO_CLASS_IDLE | 3 |
2108 +-------------------------------+---+
2110 The algorithm to set the I/O priority class for a request is as follows:
2112 - If I/O priority class policy is promote-to-rt, change the request I/O
2113 priority class to IOPRIO_CLASS_RT and change the request I/O priority
2115 - If I/O priorityt class is not promote-to-rt, translate the I/O priority
2116 class policy into a number, then change the request I/O priority class
2117 into the maximum of the I/O priority class policy number and the numerical
2123 The process number controller is used to allow a cgroup to stop any
2124 new tasks from being fork()'d or clone()'d after a specified limit is
2127 The number of tasks in a cgroup can be exhausted in ways which other
2128 controllers cannot prevent, thus warranting its own controller. For
2129 example, a fork bomb is likely to exhaust the number of tasks before
2130 hitting memory restrictions.
2132 Note that PIDs used in this controller refer to TIDs, process IDs as
2140 A read-write single value file which exists on non-root
2141 cgroups. The default is "max".
2143 Hard limit of number of processes.
2146 A read-only single value file which exists on all cgroups.
2148 The number of processes currently in the cgroup and its
2151 Organisational operations are not blocked by cgroup policies, so it is
2152 possible to have pids.current > pids.max. This can be done by either
2153 setting the limit to be smaller than pids.current, or attaching enough
2154 processes to the cgroup such that pids.current is larger than
2155 pids.max. However, it is not possible to violate a cgroup PID policy
2156 through fork() or clone(). These will return -EAGAIN if the creation
2157 of a new process would cause a cgroup policy to be violated.
2163 The "cpuset" controller provides a mechanism for constraining
2164 the CPU and memory node placement of tasks to only the resources
2165 specified in the cpuset interface files in a task's current cgroup.
2166 This is especially valuable on large NUMA systems where placing jobs
2167 on properly sized subsets of the systems with careful processor and
2168 memory placement to reduce cross-node memory access and contention
2169 can improve overall system performance.
2171 The "cpuset" controller is hierarchical. That means the controller
2172 cannot use CPUs or memory nodes not allowed in its parent.
2175 Cpuset Interface Files
2176 ~~~~~~~~~~~~~~~~~~~~~~
2179 A read-write multiple values file which exists on non-root
2180 cpuset-enabled cgroups.
2182 It lists the requested CPUs to be used by tasks within this
2183 cgroup. The actual list of CPUs to be granted, however, is
2184 subjected to constraints imposed by its parent and can differ
2185 from the requested CPUs.
2187 The CPU numbers are comma-separated numbers or ranges.
2193 An empty value indicates that the cgroup is using the same
2194 setting as the nearest cgroup ancestor with a non-empty
2195 "cpuset.cpus" or all the available CPUs if none is found.
2197 The value of "cpuset.cpus" stays constant until the next update
2198 and won't be affected by any CPU hotplug events.
2200 cpuset.cpus.effective
2201 A read-only multiple values file which exists on all
2202 cpuset-enabled cgroups.
2204 It lists the onlined CPUs that are actually granted to this
2205 cgroup by its parent. These CPUs are allowed to be used by
2206 tasks within the current cgroup.
2208 If "cpuset.cpus" is empty, the "cpuset.cpus.effective" file shows
2209 all the CPUs from the parent cgroup that can be available to
2210 be used by this cgroup. Otherwise, it should be a subset of
2211 "cpuset.cpus" unless none of the CPUs listed in "cpuset.cpus"
2212 can be granted. In this case, it will be treated just like an
2213 empty "cpuset.cpus".
2215 Its value will be affected by CPU hotplug events.
2218 A read-write multiple values file which exists on non-root
2219 cpuset-enabled cgroups.
2221 It lists the requested memory nodes to be used by tasks within
2222 this cgroup. The actual list of memory nodes granted, however,
2223 is subjected to constraints imposed by its parent and can differ
2224 from the requested memory nodes.
2226 The memory node numbers are comma-separated numbers or ranges.
2232 An empty value indicates that the cgroup is using the same
2233 setting as the nearest cgroup ancestor with a non-empty
2234 "cpuset.mems" or all the available memory nodes if none
2237 The value of "cpuset.mems" stays constant until the next update
2238 and won't be affected by any memory nodes hotplug events.
2240 Setting a non-empty value to "cpuset.mems" causes memory of
2241 tasks within the cgroup to be migrated to the designated nodes if
2242 they are currently using memory outside of the designated nodes.
2244 There is a cost for this memory migration. The migration
2245 may not be complete and some memory pages may be left behind.
2246 So it is recommended that "cpuset.mems" should be set properly
2247 before spawning new tasks into the cpuset. Even if there is
2248 a need to change "cpuset.mems" with active tasks, it shouldn't
2251 cpuset.mems.effective
2252 A read-only multiple values file which exists on all
2253 cpuset-enabled cgroups.
2255 It lists the onlined memory nodes that are actually granted to
2256 this cgroup by its parent. These memory nodes are allowed to
2257 be used by tasks within the current cgroup.
2259 If "cpuset.mems" is empty, it shows all the memory nodes from the
2260 parent cgroup that will be available to be used by this cgroup.
2261 Otherwise, it should be a subset of "cpuset.mems" unless none of
2262 the memory nodes listed in "cpuset.mems" can be granted. In this
2263 case, it will be treated just like an empty "cpuset.mems".
2265 Its value will be affected by memory nodes hotplug events.
2267 cpuset.cpus.partition
2268 A read-write single value file which exists on non-root
2269 cpuset-enabled cgroups. This flag is owned by the parent cgroup
2270 and is not delegatable.
2272 It accepts only the following input values when written to.
2274 ========== =====================================
2275 "member" Non-root member of a partition
2276 "root" Partition root
2277 "isolated" Partition root without load balancing
2278 ========== =====================================
2280 The root cgroup is always a partition root and its state
2281 cannot be changed. All other non-root cgroups start out as
2284 When set to "root", the current cgroup is the root of a new
2285 partition or scheduling domain that comprises itself and all
2286 its descendants except those that are separate partition roots
2287 themselves and their descendants.
2289 When set to "isolated", the CPUs in that partition root will
2290 be in an isolated state without any load balancing from the
2291 scheduler. Tasks placed in such a partition with multiple
2292 CPUs should be carefully distributed and bound to each of the
2293 individual CPUs for optimal performance.
2295 The value shown in "cpuset.cpus.effective" of a partition root
2296 is the CPUs that the partition root can dedicate to a potential
2297 new child partition root. The new child subtracts available
2298 CPUs from its parent "cpuset.cpus.effective".
2300 A partition root ("root" or "isolated") can be in one of the
2301 two possible states - valid or invalid. An invalid partition
2302 root is in a degraded state where some state information may
2303 be retained, but behaves more like a "member".
2305 All possible state transitions among "member", "root" and
2306 "isolated" are allowed.
2308 On read, the "cpuset.cpus.partition" file can show the following
2311 ============================= =====================================
2312 "member" Non-root member of a partition
2313 "root" Partition root
2314 "isolated" Partition root without load balancing
2315 "root invalid (<reason>)" Invalid partition root
2316 "isolated invalid (<reason>)" Invalid isolated partition root
2317 ============================= =====================================
2319 In the case of an invalid partition root, a descriptive string on
2320 why the partition is invalid is included within parentheses.
2322 For a partition root to become valid, the following conditions
2325 1) The "cpuset.cpus" is exclusive with its siblings , i.e. they
2326 are not shared by any of its siblings (exclusivity rule).
2327 2) The parent cgroup is a valid partition root.
2328 3) The "cpuset.cpus" is not empty and must contain at least
2329 one of the CPUs from parent's "cpuset.cpus", i.e. they overlap.
2330 4) The "cpuset.cpus.effective" cannot be empty unless there is
2331 no task associated with this partition.
2333 External events like hotplug or changes to "cpuset.cpus" can
2334 cause a valid partition root to become invalid and vice versa.
2335 Note that a task cannot be moved to a cgroup with empty
2336 "cpuset.cpus.effective".
2338 For a valid partition root with the sibling cpu exclusivity
2339 rule enabled, changes made to "cpuset.cpus" that violate the
2340 exclusivity rule will invalidate the partition as well as its
2341 sibling partitions with conflicting cpuset.cpus values. So
2342 care must be taking in changing "cpuset.cpus".
2344 A valid non-root parent partition may distribute out all its CPUs
2345 to its child partitions when there is no task associated with it.
2347 Care must be taken to change a valid partition root to
2348 "member" as all its child partitions, if present, will become
2349 invalid causing disruption to tasks running in those child
2350 partitions. These inactivated partitions could be recovered if
2351 their parent is switched back to a partition root with a proper
2352 set of "cpuset.cpus".
2354 Poll and inotify events are triggered whenever the state of
2355 "cpuset.cpus.partition" changes. That includes changes caused
2356 by write to "cpuset.cpus.partition", cpu hotplug or other
2357 changes that modify the validity status of the partition.
2358 This will allow user space agents to monitor unexpected changes
2359 to "cpuset.cpus.partition" without the need to do continuous
2366 Device controller manages access to device files. It includes both
2367 creation of new device files (using mknod), and access to the
2368 existing device files.
2370 Cgroup v2 device controller has no interface files and is implemented
2371 on top of cgroup BPF. To control access to device files, a user may
2372 create bpf programs of type BPF_PROG_TYPE_CGROUP_DEVICE and attach
2373 them to cgroups with BPF_CGROUP_DEVICE flag. On an attempt to access a
2374 device file, corresponding BPF programs will be executed, and depending
2375 on the return value the attempt will succeed or fail with -EPERM.
2377 A BPF_PROG_TYPE_CGROUP_DEVICE program takes a pointer to the
2378 bpf_cgroup_dev_ctx structure, which describes the device access attempt:
2379 access type (mknod/read/write) and device (type, major and minor numbers).
2380 If the program returns 0, the attempt fails with -EPERM, otherwise it
2383 An example of BPF_PROG_TYPE_CGROUP_DEVICE program may be found in
2384 tools/testing/selftests/bpf/progs/dev_cgroup.c in the kernel source tree.
2390 The "rdma" controller regulates the distribution and accounting of
2393 RDMA Interface Files
2394 ~~~~~~~~~~~~~~~~~~~~
2397 A readwrite nested-keyed file that exists for all the cgroups
2398 except root that describes current configured resource limit
2399 for a RDMA/IB device.
2401 Lines are keyed by device name and are not ordered.
2402 Each line contains space separated resource name and its configured
2403 limit that can be distributed.
2405 The following nested keys are defined.
2407 ========== =============================
2408 hca_handle Maximum number of HCA Handles
2409 hca_object Maximum number of HCA Objects
2410 ========== =============================
2412 An example for mlx4 and ocrdma device follows::
2414 mlx4_0 hca_handle=2 hca_object=2000
2415 ocrdma1 hca_handle=3 hca_object=max
2418 A read-only file that describes current resource usage.
2419 It exists for all the cgroup except root.
2421 An example for mlx4 and ocrdma device follows::
2423 mlx4_0 hca_handle=1 hca_object=20
2424 ocrdma1 hca_handle=1 hca_object=23
2429 The HugeTLB controller allows to limit the HugeTLB usage per control group and
2430 enforces the controller limit during page fault.
2432 HugeTLB Interface Files
2433 ~~~~~~~~~~~~~~~~~~~~~~~
2435 hugetlb.<hugepagesize>.current
2436 Show current usage for "hugepagesize" hugetlb. It exists for all
2437 the cgroup except root.
2439 hugetlb.<hugepagesize>.max
2440 Set/show the hard limit of "hugepagesize" hugetlb usage.
2441 The default value is "max". It exists for all the cgroup except root.
2443 hugetlb.<hugepagesize>.events
2444 A read-only flat-keyed file which exists on non-root cgroups.
2447 The number of allocation failure due to HugeTLB limit
2449 hugetlb.<hugepagesize>.events.local
2450 Similar to hugetlb.<hugepagesize>.events but the fields in the file
2451 are local to the cgroup i.e. not hierarchical. The file modified event
2452 generated on this file reflects only the local events.
2454 hugetlb.<hugepagesize>.numa_stat
2455 Similar to memory.numa_stat, it shows the numa information of the
2456 hugetlb pages of <hugepagesize> in this cgroup. Only active in
2457 use hugetlb pages are included. The per-node values are in bytes.
2462 The Miscellaneous cgroup provides the resource limiting and tracking
2463 mechanism for the scalar resources which cannot be abstracted like the other
2464 cgroup resources. Controller is enabled by the CONFIG_CGROUP_MISC config
2467 A resource can be added to the controller via enum misc_res_type{} in the
2468 include/linux/misc_cgroup.h file and the corresponding name via misc_res_name[]
2469 in the kernel/cgroup/misc.c file. Provider of the resource must set its
2470 capacity prior to using the resource by calling misc_cg_set_capacity().
2472 Once a capacity is set then the resource usage can be updated using charge and
2473 uncharge APIs. All of the APIs to interact with misc controller are in
2474 include/linux/misc_cgroup.h.
2476 Misc Interface Files
2477 ~~~~~~~~~~~~~~~~~~~~
2479 Miscellaneous controller provides 3 interface files. If two misc resources (res_a and res_b) are registered then:
2482 A read-only flat-keyed file shown only in the root cgroup. It shows
2483 miscellaneous scalar resources available on the platform along with
2491 A read-only flat-keyed file shown in the all cgroups. It shows
2492 the current usage of the resources in the cgroup and its children.::
2499 A read-write flat-keyed file shown in the non root cgroups. Allowed
2500 maximum usage of the resources in the cgroup and its children.::
2506 Limit can be set by::
2508 # echo res_a 1 > misc.max
2510 Limit can be set to max by::
2512 # echo res_a max > misc.max
2514 Limits can be set higher than the capacity value in the misc.capacity
2518 A read-only flat-keyed file which exists on non-root cgroups. The
2519 following entries are defined. Unless specified otherwise, a value
2520 change in this file generates a file modified event. All fields in
2521 this file are hierarchical.
2524 The number of times the cgroup's resource usage was
2525 about to go over the max boundary.
2527 Migration and Ownership
2528 ~~~~~~~~~~~~~~~~~~~~~~~
2530 A miscellaneous scalar resource is charged to the cgroup in which it is used
2531 first, and stays charged to that cgroup until that resource is freed. Migrating
2532 a process to a different cgroup does not move the charge to the destination
2533 cgroup where the process has moved.
2541 perf_event controller, if not mounted on a legacy hierarchy, is
2542 automatically enabled on the v2 hierarchy so that perf events can
2543 always be filtered by cgroup v2 path. The controller can still be
2544 moved to a legacy hierarchy after v2 hierarchy is populated.
2547 Non-normative information
2548 -------------------------
2550 This section contains information that isn't considered to be a part of
2551 the stable kernel API and so is subject to change.
2554 CPU controller root cgroup process behaviour
2555 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2557 When distributing CPU cycles in the root cgroup each thread in this
2558 cgroup is treated as if it was hosted in a separate child cgroup of the
2559 root cgroup. This child cgroup weight is dependent on its thread nice
2562 For details of this mapping see sched_prio_to_weight array in
2563 kernel/sched/core.c file (values from this array should be scaled
2564 appropriately so the neutral - nice 0 - value is 100 instead of 1024).
2567 IO controller root cgroup process behaviour
2568 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2570 Root cgroup processes are hosted in an implicit leaf child node.
2571 When distributing IO resources this implicit child node is taken into
2572 account as if it was a normal child cgroup of the root cgroup with a
2573 weight value of 200.
2582 cgroup namespace provides a mechanism to virtualize the view of the
2583 "/proc/$PID/cgroup" file and cgroup mounts. The CLONE_NEWCGROUP clone
2584 flag can be used with clone(2) and unshare(2) to create a new cgroup
2585 namespace. The process running inside the cgroup namespace will have
2586 its "/proc/$PID/cgroup" output restricted to cgroupns root. The
2587 cgroupns root is the cgroup of the process at the time of creation of
2588 the cgroup namespace.
2590 Without cgroup namespace, the "/proc/$PID/cgroup" file shows the
2591 complete path of the cgroup of a process. In a container setup where
2592 a set of cgroups and namespaces are intended to isolate processes the
2593 "/proc/$PID/cgroup" file may leak potential system level information
2594 to the isolated processes. For example::
2596 # cat /proc/self/cgroup
2597 0::/batchjobs/container_id1
2599 The path '/batchjobs/container_id1' can be considered as system-data
2600 and undesirable to expose to the isolated processes. cgroup namespace
2601 can be used to restrict visibility of this path. For example, before
2602 creating a cgroup namespace, one would see::
2604 # ls -l /proc/self/ns/cgroup
2605 lrwxrwxrwx 1 root root 0 2014-07-15 10:37 /proc/self/ns/cgroup -> cgroup:[4026531835]
2606 # cat /proc/self/cgroup
2607 0::/batchjobs/container_id1
2609 After unsharing a new namespace, the view changes::
2611 # ls -l /proc/self/ns/cgroup
2612 lrwxrwxrwx 1 root root 0 2014-07-15 10:35 /proc/self/ns/cgroup -> cgroup:[4026532183]
2613 # cat /proc/self/cgroup
2616 When some thread from a multi-threaded process unshares its cgroup
2617 namespace, the new cgroupns gets applied to the entire process (all
2618 the threads). This is natural for the v2 hierarchy; however, for the
2619 legacy hierarchies, this may be unexpected.
2621 A cgroup namespace is alive as long as there are processes inside or
2622 mounts pinning it. When the last usage goes away, the cgroup
2623 namespace is destroyed. The cgroupns root and the actual cgroups
2630 The 'cgroupns root' for a cgroup namespace is the cgroup in which the
2631 process calling unshare(2) is running. For example, if a process in
2632 /batchjobs/container_id1 cgroup calls unshare, cgroup
2633 /batchjobs/container_id1 becomes the cgroupns root. For the
2634 init_cgroup_ns, this is the real root ('/') cgroup.
2636 The cgroupns root cgroup does not change even if the namespace creator
2637 process later moves to a different cgroup::
2639 # ~/unshare -c # unshare cgroupns in some cgroup
2640 # cat /proc/self/cgroup
2643 # echo 0 > sub_cgrp_1/cgroup.procs
2644 # cat /proc/self/cgroup
2647 Each process gets its namespace-specific view of "/proc/$PID/cgroup"
2649 Processes running inside the cgroup namespace will be able to see
2650 cgroup paths (in /proc/self/cgroup) only inside their root cgroup.
2651 From within an unshared cgroupns::
2655 # echo 7353 > sub_cgrp_1/cgroup.procs
2656 # cat /proc/7353/cgroup
2659 From the initial cgroup namespace, the real cgroup path will be
2662 $ cat /proc/7353/cgroup
2663 0::/batchjobs/container_id1/sub_cgrp_1
2665 From a sibling cgroup namespace (that is, a namespace rooted at a
2666 different cgroup), the cgroup path relative to its own cgroup
2667 namespace root will be shown. For instance, if PID 7353's cgroup
2668 namespace root is at '/batchjobs/container_id2', then it will see::
2670 # cat /proc/7353/cgroup
2671 0::/../container_id2/sub_cgrp_1
2673 Note that the relative path always starts with '/' to indicate that
2674 its relative to the cgroup namespace root of the caller.
2677 Migration and setns(2)
2678 ----------------------
2680 Processes inside a cgroup namespace can move into and out of the
2681 namespace root if they have proper access to external cgroups. For
2682 example, from inside a namespace with cgroupns root at
2683 /batchjobs/container_id1, and assuming that the global hierarchy is
2684 still accessible inside cgroupns::
2686 # cat /proc/7353/cgroup
2688 # echo 7353 > batchjobs/container_id2/cgroup.procs
2689 # cat /proc/7353/cgroup
2690 0::/../container_id2
2692 Note that this kind of setup is not encouraged. A task inside cgroup
2693 namespace should only be exposed to its own cgroupns hierarchy.
2695 setns(2) to another cgroup namespace is allowed when:
2697 (a) the process has CAP_SYS_ADMIN against its current user namespace
2698 (b) the process has CAP_SYS_ADMIN against the target cgroup
2701 No implicit cgroup changes happen with attaching to another cgroup
2702 namespace. It is expected that the someone moves the attaching
2703 process under the target cgroup namespace root.
2706 Interaction with Other Namespaces
2707 ---------------------------------
2709 Namespace specific cgroup hierarchy can be mounted by a process
2710 running inside a non-init cgroup namespace::
2712 # mount -t cgroup2 none $MOUNT_POINT
2714 This will mount the unified cgroup hierarchy with cgroupns root as the
2715 filesystem root. The process needs CAP_SYS_ADMIN against its user and
2718 The virtualization of /proc/self/cgroup file combined with restricting
2719 the view of cgroup hierarchy by namespace-private cgroupfs mount
2720 provides a properly isolated cgroup view inside the container.
2723 Information on Kernel Programming
2724 =================================
2726 This section contains kernel programming information in the areas
2727 where interacting with cgroup is necessary. cgroup core and
2728 controllers are not covered.
2731 Filesystem Support for Writeback
2732 --------------------------------
2734 A filesystem can support cgroup writeback by updating
2735 address_space_operations->writepage[s]() to annotate bio's using the
2736 following two functions.
2738 wbc_init_bio(@wbc, @bio)
2739 Should be called for each bio carrying writeback data and
2740 associates the bio with the inode's owner cgroup and the
2741 corresponding request queue. This must be called after
2742 a queue (device) has been associated with the bio and
2745 wbc_account_cgroup_owner(@wbc, @page, @bytes)
2746 Should be called for each data segment being written out.
2747 While this function doesn't care exactly when it's called
2748 during the writeback session, it's the easiest and most
2749 natural to call it as data segments are added to a bio.
2751 With writeback bio's annotated, cgroup support can be enabled per
2752 super_block by setting SB_I_CGROUPWB in ->s_iflags. This allows for
2753 selective disabling of cgroup writeback support which is helpful when
2754 certain filesystem features, e.g. journaled data mode, are
2757 wbc_init_bio() binds the specified bio to its cgroup. Depending on
2758 the configuration, the bio may be executed at a lower priority and if
2759 the writeback session is holding shared resources, e.g. a journal
2760 entry, may lead to priority inversion. There is no one easy solution
2761 for the problem. Filesystems can try to work around specific problem
2762 cases by skipping wbc_init_bio() and using bio_associate_blkg()
2766 Deprecated v1 Core Features
2767 ===========================
2769 - Multiple hierarchies including named ones are not supported.
2771 - All v1 mount options are not supported.
2773 - The "tasks" file is removed and "cgroup.procs" is not sorted.
2775 - "cgroup.clone_children" is removed.
2777 - /proc/cgroups is meaningless for v2. Use "cgroup.controllers" file
2778 at the root instead.
2781 Issues with v1 and Rationales for v2
2782 ====================================
2784 Multiple Hierarchies
2785 --------------------
2787 cgroup v1 allowed an arbitrary number of hierarchies and each
2788 hierarchy could host any number of controllers. While this seemed to
2789 provide a high level of flexibility, it wasn't useful in practice.
2791 For example, as there is only one instance of each controller, utility
2792 type controllers such as freezer which can be useful in all
2793 hierarchies could only be used in one. The issue is exacerbated by
2794 the fact that controllers couldn't be moved to another hierarchy once
2795 hierarchies were populated. Another issue was that all controllers
2796 bound to a hierarchy were forced to have exactly the same view of the
2797 hierarchy. It wasn't possible to vary the granularity depending on
2798 the specific controller.
2800 In practice, these issues heavily limited which controllers could be
2801 put on the same hierarchy and most configurations resorted to putting
2802 each controller on its own hierarchy. Only closely related ones, such
2803 as the cpu and cpuacct controllers, made sense to be put on the same
2804 hierarchy. This often meant that userland ended up managing multiple
2805 similar hierarchies repeating the same steps on each hierarchy
2806 whenever a hierarchy management operation was necessary.
2808 Furthermore, support for multiple hierarchies came at a steep cost.
2809 It greatly complicated cgroup core implementation but more importantly
2810 the support for multiple hierarchies restricted how cgroup could be
2811 used in general and what controllers was able to do.
2813 There was no limit on how many hierarchies there might be, which meant
2814 that a thread's cgroup membership couldn't be described in finite
2815 length. The key might contain any number of entries and was unlimited
2816 in length, which made it highly awkward to manipulate and led to
2817 addition of controllers which existed only to identify membership,
2818 which in turn exacerbated the original problem of proliferating number
2821 Also, as a controller couldn't have any expectation regarding the
2822 topologies of hierarchies other controllers might be on, each
2823 controller had to assume that all other controllers were attached to
2824 completely orthogonal hierarchies. This made it impossible, or at
2825 least very cumbersome, for controllers to cooperate with each other.
2827 In most use cases, putting controllers on hierarchies which are
2828 completely orthogonal to each other isn't necessary. What usually is
2829 called for is the ability to have differing levels of granularity
2830 depending on the specific controller. In other words, hierarchy may
2831 be collapsed from leaf towards root when viewed from specific
2832 controllers. For example, a given configuration might not care about
2833 how memory is distributed beyond a certain level while still wanting
2834 to control how CPU cycles are distributed.
2840 cgroup v1 allowed threads of a process to belong to different cgroups.
2841 This didn't make sense for some controllers and those controllers
2842 ended up implementing different ways to ignore such situations but
2843 much more importantly it blurred the line between API exposed to
2844 individual applications and system management interface.
2846 Generally, in-process knowledge is available only to the process
2847 itself; thus, unlike service-level organization of processes,
2848 categorizing threads of a process requires active participation from
2849 the application which owns the target process.
2851 cgroup v1 had an ambiguously defined delegation model which got abused
2852 in combination with thread granularity. cgroups were delegated to
2853 individual applications so that they can create and manage their own
2854 sub-hierarchies and control resource distributions along them. This
2855 effectively raised cgroup to the status of a syscall-like API exposed
2858 First of all, cgroup has a fundamentally inadequate interface to be
2859 exposed this way. For a process to access its own knobs, it has to
2860 extract the path on the target hierarchy from /proc/self/cgroup,
2861 construct the path by appending the name of the knob to the path, open
2862 and then read and/or write to it. This is not only extremely clunky
2863 and unusual but also inherently racy. There is no conventional way to
2864 define transaction across the required steps and nothing can guarantee
2865 that the process would actually be operating on its own sub-hierarchy.
2867 cgroup controllers implemented a number of knobs which would never be
2868 accepted as public APIs because they were just adding control knobs to
2869 system-management pseudo filesystem. cgroup ended up with interface
2870 knobs which were not properly abstracted or refined and directly
2871 revealed kernel internal details. These knobs got exposed to
2872 individual applications through the ill-defined delegation mechanism
2873 effectively abusing cgroup as a shortcut to implementing public APIs
2874 without going through the required scrutiny.
2876 This was painful for both userland and kernel. Userland ended up with
2877 misbehaving and poorly abstracted interfaces and kernel exposing and
2878 locked into constructs inadvertently.
2881 Competition Between Inner Nodes and Threads
2882 -------------------------------------------
2884 cgroup v1 allowed threads to be in any cgroups which created an
2885 interesting problem where threads belonging to a parent cgroup and its
2886 children cgroups competed for resources. This was nasty as two
2887 different types of entities competed and there was no obvious way to
2888 settle it. Different controllers did different things.
2890 The cpu controller considered threads and cgroups as equivalents and
2891 mapped nice levels to cgroup weights. This worked for some cases but
2892 fell flat when children wanted to be allocated specific ratios of CPU
2893 cycles and the number of internal threads fluctuated - the ratios
2894 constantly changed as the number of competing entities fluctuated.
2895 There also were other issues. The mapping from nice level to weight
2896 wasn't obvious or universal, and there were various other knobs which
2897 simply weren't available for threads.
2899 The io controller implicitly created a hidden leaf node for each
2900 cgroup to host the threads. The hidden leaf had its own copies of all
2901 the knobs with ``leaf_`` prefixed. While this allowed equivalent
2902 control over internal threads, it was with serious drawbacks. It
2903 always added an extra layer of nesting which wouldn't be necessary
2904 otherwise, made the interface messy and significantly complicated the
2907 The memory controller didn't have a way to control what happened
2908 between internal tasks and child cgroups and the behavior was not
2909 clearly defined. There were attempts to add ad-hoc behaviors and
2910 knobs to tailor the behavior to specific workloads which would have
2911 led to problems extremely difficult to resolve in the long term.
2913 Multiple controllers struggled with internal tasks and came up with
2914 different ways to deal with it; unfortunately, all the approaches were
2915 severely flawed and, furthermore, the widely different behaviors
2916 made cgroup as a whole highly inconsistent.
2918 This clearly is a problem which needs to be addressed from cgroup core
2922 Other Interface Issues
2923 ----------------------
2925 cgroup v1 grew without oversight and developed a large number of
2926 idiosyncrasies and inconsistencies. One issue on the cgroup core side
2927 was how an empty cgroup was notified - a userland helper binary was
2928 forked and executed for each event. The event delivery wasn't
2929 recursive or delegatable. The limitations of the mechanism also led
2930 to in-kernel event delivery filtering mechanism further complicating
2933 Controller interfaces were problematic too. An extreme example is
2934 controllers completely ignoring hierarchical organization and treating
2935 all cgroups as if they were all located directly under the root
2936 cgroup. Some controllers exposed a large amount of inconsistent
2937 implementation details to userland.
2939 There also was no consistency across controllers. When a new cgroup
2940 was created, some controllers defaulted to not imposing extra
2941 restrictions while others disallowed any resource usage until
2942 explicitly configured. Configuration knobs for the same type of
2943 control used widely differing naming schemes and formats. Statistics
2944 and information knobs were named arbitrarily and used different
2945 formats and units even in the same controller.
2947 cgroup v2 establishes common conventions where appropriate and updates
2948 controllers so that they expose minimal and consistent interfaces.
2951 Controller Issues and Remedies
2952 ------------------------------
2957 The original lower boundary, the soft limit, is defined as a limit
2958 that is per default unset. As a result, the set of cgroups that
2959 global reclaim prefers is opt-in, rather than opt-out. The costs for
2960 optimizing these mostly negative lookups are so high that the
2961 implementation, despite its enormous size, does not even provide the
2962 basic desirable behavior. First off, the soft limit has no
2963 hierarchical meaning. All configured groups are organized in a global
2964 rbtree and treated like equal peers, regardless where they are located
2965 in the hierarchy. This makes subtree delegation impossible. Second,
2966 the soft limit reclaim pass is so aggressive that it not just
2967 introduces high allocation latencies into the system, but also impacts
2968 system performance due to overreclaim, to the point where the feature
2969 becomes self-defeating.
2971 The memory.low boundary on the other hand is a top-down allocated
2972 reserve. A cgroup enjoys reclaim protection when it's within its
2973 effective low, which makes delegation of subtrees possible. It also
2974 enjoys having reclaim pressure proportional to its overage when
2975 above its effective low.
2977 The original high boundary, the hard limit, is defined as a strict
2978 limit that can not budge, even if the OOM killer has to be called.
2979 But this generally goes against the goal of making the most out of the
2980 available memory. The memory consumption of workloads varies during
2981 runtime, and that requires users to overcommit. But doing that with a
2982 strict upper limit requires either a fairly accurate prediction of the
2983 working set size or adding slack to the limit. Since working set size
2984 estimation is hard and error prone, and getting it wrong results in
2985 OOM kills, most users tend to err on the side of a looser limit and
2986 end up wasting precious resources.
2988 The memory.high boundary on the other hand can be set much more
2989 conservatively. When hit, it throttles allocations by forcing them
2990 into direct reclaim to work off the excess, but it never invokes the
2991 OOM killer. As a result, a high boundary that is chosen too
2992 aggressively will not terminate the processes, but instead it will
2993 lead to gradual performance degradation. The user can monitor this
2994 and make corrections until the minimal memory footprint that still
2995 gives acceptable performance is found.
2997 In extreme cases, with many concurrent allocations and a complete
2998 breakdown of reclaim progress within the group, the high boundary can
2999 be exceeded. But even then it's mostly better to satisfy the
3000 allocation from the slack available in other groups or the rest of the
3001 system than killing the group. Otherwise, memory.max is there to
3002 limit this type of spillover and ultimately contain buggy or even
3003 malicious applications.
3005 Setting the original memory.limit_in_bytes below the current usage was
3006 subject to a race condition, where concurrent charges could cause the
3007 limit setting to fail. memory.max on the other hand will first set the
3008 limit to prevent new charges, and then reclaim and OOM kill until the
3009 new limit is met - or the task writing to memory.max is killed.
3011 The combined memory+swap accounting and limiting is replaced by real
3012 control over swap space.
3014 The main argument for a combined memory+swap facility in the original
3015 cgroup design was that global or parental pressure would always be
3016 able to swap all anonymous memory of a child group, regardless of the
3017 child's own (possibly untrusted) configuration. However, untrusted
3018 groups can sabotage swapping by other means - such as referencing its
3019 anonymous memory in a tight loop - and an admin can not assume full
3020 swappability when overcommitting untrusted jobs.
3022 For trusted jobs, on the other hand, a combined counter is not an
3023 intuitive userspace interface, and it flies in the face of the idea
3024 that cgroup controllers should account and limit specific physical
3025 resources. Swap space is a resource like all others in the system,
3026 and that's why unified hierarchy allows distributing it separately.